Quality-Informed Active Learning on Social Media in Crisis Scenarios

During large-scale emergencies, timely situational awareness is crucial for effective response actions aimed at mitigating losses. With their real-time and distributed nature, social media provide valuable support for situational awareness. However, the unpredictability of emergency scenarios poses a challenge in understanding which data are relevant, especially in the early stages. This study explores the challenge of learning to identify relevant data dynamically as critical events unfold, operating with limited ground-truth availability in an Active Learning paradigm. We propose a tailored framework based on Information Quality indicators and evaluate it across multiple crisis events. We assess the importance of the proposed indicators for characterizing social media posts across different events, notions of relevance, and temporal intervals. We then explore the usability of these indicators to perform Active Learning in evolving emergency scenarios. Finally, we investigate whether the importance of individual posts can be generalized. Based on the insights obtained on the role of Information Quality, the steps to advance this line of research are outlined.